Supplementary Materials

I have set echo=FALSE so that most of the code chunks will not display. Please refer to the .Rmd file for source code.

1 Loading packages and importing data

1.1 Packages

Details of packages, etc, are given below.

## R version 3.6.1 (2019-07-05)
## Platform: x86_64-apple-darwin15.6.0 (64-bit)
## Running under: macOS Mojave 10.14.6
## 
## Matrix products: default
## BLAS:   /Library/Frameworks/R.framework/Versions/3.6/Resources/lib/libRblas.0.dylib
## LAPACK: /Library/Frameworks/R.framework/Versions/3.6/Resources/lib/libRlapack.dylib
## 
## locale:
## [1] en_GB.UTF-8/en_GB.UTF-8/en_GB.UTF-8/C/en_GB.UTF-8/en_GB.UTF-8
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
##  [1] magick_2.3      patchwork_1.0.0 ggthemes_4.2.0  see_0.2.1      
##  [5] latex2exp_0.4.0 ggridges_0.5.1  tidybayes_2.0.2 brms_2.12.0    
##  [9] Rcpp_1.0.3      forcats_0.4.0   stringr_1.4.0   dplyr_0.8.3    
## [13] purrr_0.3.3     readr_1.3.1     tidyr_1.0.2     tibble_2.1.3   
## [17] ggplot2_3.2.1   tidyverse_1.2.1 bookdown_0.18  
## 
## loaded via a namespace (and not attached):
##  [1] colorspace_1.4-1          rsconnect_0.8.15         
##  [3] markdown_1.1              base64enc_0.1-3          
##  [5] rstudioapi_0.10           rstan_2.19.2             
##  [7] svUnit_0.7-12             DT_0.8                   
##  [9] fansi_0.4.1               lubridate_1.7.4          
## [11] xml2_1.2.2                bridgesampling_0.7-2     
## [13] knitr_1.25                shinythemes_1.1.2        
## [15] bayesplot_1.7.0           jsonlite_1.6             
## [17] broom_0.5.2               shiny_1.3.2              
## [19] compiler_3.6.1            httr_1.4.1               
## [21] backports_1.1.5           assertthat_0.2.1         
## [23] Matrix_1.2-17             lazyeval_0.2.2           
## [25] cli_2.0.1                 later_0.8.0              
## [27] htmltools_0.3.6           prettyunits_1.1.1        
## [29] tools_3.6.1               igraph_1.2.4.1           
## [31] coda_0.19-3               gtable_0.3.0             
## [33] glue_1.3.1                reshape2_1.4.3           
## [35] cellranger_1.1.0          vctrs_0.2.2              
## [37] nlme_3.1-140              crosstalk_1.0.0          
## [39] insight_0.5.0             xfun_0.8                 
## [41] ps_1.3.0                  rvest_0.3.4              
## [43] mime_0.7                  miniUI_0.1.1.1           
## [45] lifecycle_0.1.0           gtools_3.8.1             
## [47] zoo_1.8-6                 scales_1.1.0             
## [49] colourpicker_1.0          hms_0.5.0                
## [51] promises_1.0.1            Brobdingnag_1.2-6        
## [53] parallel_3.6.1            inline_0.3.15            
## [55] shinystan_2.5.0           yaml_2.2.0               
## [57] gridExtra_2.3             loo_2.2.0                
## [59] StanHeaders_2.19.0        stringi_1.4.5            
## [61] bayestestR_0.3.0          dygraphs_1.1.1.6         
## [63] pkgbuild_1.0.6            rlang_0.4.4              
## [65] pkgconfig_2.0.3           matrixStats_0.55.0       
## [67] evaluate_0.14             lattice_0.20-38          
## [69] rstantools_2.0.0          htmlwidgets_1.3          
## [71] tidyselect_0.2.5          processx_3.4.1           
## [73] plyr_1.8.5                magrittr_1.5             
## [75] R6_2.4.1                  generics_0.0.2           
## [77] pillar_1.4.3              haven_2.1.1              
## [79] withr_2.1.2               xts_0.11-2               
## [81] abind_1.4-5               modelr_0.1.5             
## [83] crayon_1.3.4              arrayhelpers_1.0-20160527
## [85] rmarkdown_2.1             grid_3.6.1               
## [87] readxl_1.3.1              callr_3.4.1              
## [89] threejs_0.3.1             digest_0.6.23            
## [91] xtable_1.8-4              httpuv_1.5.1             
## [93] stats4_3.6.1              munsell_0.5.0            
## [95] shinyjs_1.0

1.2 Import Data

Here we import our data and make some summary plots.

1.2.1 EEG Data

Importing the primary EEG data set. This is odd-harmonic filtered data from region-of-interest consisting of six electrodes over occipital cortex.

Figure 1.1: RMS data from two of the wallpaper groups, P2 and P4M. Odd harmonics are shown in A and B, while even harmonic data are shown in C and D, and occipital and parietal regions of interest are shown in dark nad light gray, respectively. The two groups elicit very different response amplitudes for odd harmonics over occipital cortex, but for even harmonics those differences are much less pronounced.

## Observations: 400
## Variables: 3
## $ wg      <chr> "P2", "PM", "PG", "CM", "PMM", "PMG", "PGG", "CMM", "P4"…
## $ subject <chr> "s01", "s01", "s01", "s01", "s01", "s01", "s01", "s01", …
## $ rms     <dbl> 0.4013, 0.6555, 0.5547, 0.7635, 0.9185, 0.7285, 0.4320, …

It’s clearly skewed, and negative display duration are impossible, so will fit a glm with family = 'lognormal'.

Figure 1.2: Data are the root-mean-squaured (rms) over the odd-harmonic filtered waveforms.

1.2.2 Threshold Data

Here we import that data and select the columns that we’re interested in. Threshold gives the required display duration (in seconds) for the two stimuli to allow for accurate discrimination.

## Observations: 186
## Variables: 3
## $ subject   <chr> "person10", "person10", "person10", "person10", "perso…
## $ wg        <chr> "CM", "CMM", "P2", "P3", "P31M", "P3M1", "P4", "P4G", …
## $ threshold <dbl> 0.74125, 0.20216, 0.47697, 0.35012, 0.24529, 0.19022, …

As above, a summary of the data.

Figure 1.3: Histogram of the display duration thresholds.

Again, we have a skewed distribution, so will fit with family = 'lognormal'.

1.2.3 Control Data

In addition to the primary EEG data set, we are also importing two control data sets which are (a) even harmonic data from the same occipital electrodes, and (b) odd harmonic data from six parietal electrodes (see Figure 1.1 and the main paper).

1.3 Subgroup Information

Import subgroup information and display a table of the relationships that we will be investigating.

## Observations: 69
## Variables: 5
## $ group    <chr> "PG", "CM", "PG", "PM", "P2", "PG", "P2", "PG", "PM", "…
## $ subgroup <chr> "PM", "PM", "CM", "CM", "PGG", "PGG", "PMG", "PMG", "PM…
## $ index    <dbl> 2, 2, 2, 2, 2, 2, 2, 2, 2, 4, 2, 2, 4, 2, 4, 4, 2, 2, 2…
## $ normal   <lgl> TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, TRUE, F…
## $ label    <chr> "PG $\\rightarrow$ PM", "CM $\\rightarrow$ PM", "PG $\\…

Table 1.1: subgroup index

subgroup	P2	PG	PM	CM	PGG	PMG	PMM	CMM	P4	P4G	P4M	P3	P3M1	P31M	P6	P6M
P2	2
PG		2
PM		2	2	2
CM		2	2	3
PGG	2	2			3
PMG	2	2	2	4	2	3
PMM	2	4	2	4	4	2	2	2
CMM	2	4	4	2	2	2	2	4
P4	2								2
P4G	4	4	8	4	2	4	4	2	2	9
P4M	4	8	4	4	4	4	2	2	2	2	2
P3												3
P3M1		6	6	3								2	4	3
P31M		6	6	3								2	3	4
P6	3											2			4
P6M	6	12	12	6	6	6	6	3				4	2	2	2	3

Table 1.2: normal and non-normal subgroups

subgroup	P2	PG	PM	CM	PGG	PMG	PMM	CMM	P4	P4G	P4M	P3	P3M1	P31M	P6	P6M
P2	TRUE
PG		TRUE
PM		TRUE	TRUE	TRUE
CM		TRUE	TRUE	FALSE
PGG	TRUE	TRUE			FALSE
PMG	TRUE	TRUE	TRUE	FALSE	TRUE	FALSE
PMM	TRUE	TRUE	TRUE	TRUE	TRUE	TRUE	TRUE	TRUE
CMM	TRUE	TRUE	TRUE	TRUE	TRUE	TRUE	TRUE	FALSE
P4	TRUE								TRUE
P4G	TRUE	FALSE	FALSE	FALSE	TRUE	FALSE	TRUE	TRUE	TRUE	FALSE
P4M	TRUE	FALSE	FALSE	FALSE	TRUE	FALSE	TRUE	TRUE	TRUE	TRUE	TRUE
P3												TRUE
P3M1		FALSE	FALSE	FALSE								TRUE	FALSE	TRUE
P31M		FALSE	FALSE	FALSE								TRUE	FALSE	FALSE
P6	TRUE											TRUE			FALSE
P6M	TRUE	FALSE	FALSE	FALSE	FALSE	FALSE	FALSE	FALSE				TRUE	TRUE	TRUE	TRUE	FALSE

2 Bayesian Analysis

Here are the details of the Bayesian multi-level modelling. Our general approach is:

• Use weakly-informative priors.
• Fit a multi-level model using the lognormal distribution.
• independent variable (wallpaper group) and one dependent variable (rms or threshold).
• maximal random effect structure.
• Check model fit (r̂ statistics and trace plots).
• Extract mcmc samples from the posterior distributions.
• Use these samples to estimate the difference between sub- and super-groups.

2.1 Define Priors

In this section we will specify some priors. We then then use a prior-predictive check to assess whether the prior is reasonable or not (i.e., on the same order of magnitude as our measurements).

2.1.1 Fixed Effects

Our independent variable is a categorical factor with 16 levels. We will drop the intercept from our model and instead fit a coefficent for each factor level (\(y \sim x - 0\)). As our dependant variable will be log-transformed, we can use the priors below:

prior <- c(
  set_prior("normal(0,2)", class = "b"),    
  set_prior("cauchy(0,2)", class = "sigma"))

2.1.2 Group-level Effects

We will keep the default weakly informative priors for the group-level (‘random’) effects. From the brms documentation:

[…] restricted to be non-negative and, by default, have a half student-t prior with 3 degrees of freedom and a scale parameter that depends on the standard deviation of the response after applying the link function. Minimally, the scale parameter is 10. This prior is used (a) to be only very weakly informative in order to influence results as few as possible, while (b) providing at least some regularization to considerably improve convergence and sampling efficiency.

2.1.3 Prior Predictive Check

Now we can specify our Bayesian multi-level model and priors. Note that as we are using sample_prior = 'only', the model will not learn anything from our data.

m_prior <- brm(data = d_eeg, 
  rms ~ wg-1 + (1|subject),
  family = "lognormal", 
  prior = prior, 
  iter = n_iter ,
  sample_prior = 'only')

We can use this model to generate data.

## Observations: 320,000
## Variables: 2
## $ key   <chr> "P2", "P2", "P2", "P2", "P2", "P2", "P2", "P2", "P2", "P2"…
## $ value <dbl> 2.58726956, 0.39915973, 0.60307342, 0.70885249, 0.11650792…

Figure 2.1: The density plot shows the distribution of the empirical data, while the blue line shows the 66% and 95% prediction intervals.

We can see that i) our priors are relatively weak as the predictions span several orders of magnitide, and ii) our empirical data falls within this range.

2.2 Compute Posterior

2.2.1 Fit Model to EEG Data

We will now fit the model to the data.

##  Family: lognormal 
##   Links: mu = identity; sigma = identity 
## Formula: rms ~ wg - 1 + (1 | subject) 
##    Data: d_eeg (Number of observations: 400) 
## Samples: 4 chains, each with iter = 10000; warmup = 5000; thin = 1;
##          total post-warmup samples = 20000
## 
## Group-Level Effects: 
## ~subject (Number of levels: 25) 
##               Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sd(Intercept)     0.38      0.06     0.28     0.52 1.00     1417     2504
## 
## Population-Level Effects: 
##        Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## wgCM      -0.49      0.09    -0.66    -0.31 1.00      877     1502
## wgCMM     -0.13      0.09    -0.29     0.05 1.00      845     1513
## wgP2      -0.95      0.09    -1.12    -0.77 1.00      873     1536
## wgP3      -0.82      0.09    -0.99    -0.64 1.01      854     1517
## wgP31M    -0.44      0.09    -0.61    -0.26 1.00      861     1337
## wgP3M1    -0.14      0.09    -0.31     0.04 1.00      870     1559
## wgP4      -0.48      0.09    -0.66    -0.31 1.00      884     1412
## wgP4G     -0.26      0.09    -0.43    -0.08 1.00      842     1483
## wgP4M      0.15      0.09    -0.02     0.33 1.00      896     1618
## wgP6      -0.48      0.09    -0.66    -0.31 1.00      847     1536
## wgP6M     -0.05      0.09    -0.22     0.13 1.00      860     1591
## wgPG      -0.93      0.09    -1.11    -0.75 1.00      849     1471
## wgPGG     -0.73      0.09    -0.90    -0.55 1.00      851     1504
## wgPM      -0.59      0.09    -0.77    -0.42 1.00      852     1373
## wgPMG     -0.26      0.09    -0.43    -0.08 1.00      847     1549
## wgPMM     -0.07      0.09    -0.24     0.11 1.01      860     1378
## 
## Family Specific Parameters: 
##       Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sigma     0.23      0.01     0.21     0.25 1.00     8172    10803
## 
## Samples were drawn using sampling(NUTS). For each parameter, Bulk_ESS
## and Tail_ESS are effective sample size measures, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).

We will now look at the model’s predicts for the average participant (i.e, ignoring the random intercepts).

Figure 2.2: The density plot shows the distribution of the empirical data, while the blue line shows the 66% and 95% prediction intervals.

2.2.2 Fit Model to Psychophysical Data

We will now fit the model to the data.

##  Family: lognormal 
##   Links: mu = identity; sigma = identity 
## Formula: threshold ~ wg - 1 + (1 | subject) 
##    Data: d_dispthresh (Number of observations: 186) 
## Samples: 4 chains, each with iter = 10000; warmup = 5000; thin = 1;
##          total post-warmup samples = 20000
## 
## Group-Level Effects: 
## ~subject (Number of levels: 12) 
##               Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sd(Intercept)     0.39      0.11     0.23     0.66 1.00     3568     5476
## 
## Population-Level Effects: 
##        Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## wgCM      -0.22      0.17    -0.53     0.12 1.00     2280     3879
## wgCMM     -1.15      0.17    -1.47    -0.81 1.00     2259     3668
## wgP2      -0.29      0.17    -0.61     0.06 1.00     2578     4062
## wgP3      -0.69      0.17    -1.00    -0.35 1.00     2289     3760
## wgP31M    -1.18      0.17    -1.49    -0.85 1.00     2338     4052
## wgP3M1    -1.35      0.17    -1.66    -1.01 1.00     2273     3518
## wgP4      -0.89      0.17    -1.21    -0.55 1.00     2361     3303
## wgP4G     -1.22      0.17    -1.54    -0.88 1.00     2484     4249
## wgP4M     -1.29      0.17    -1.61    -0.95 1.00     2181     3222
## wgP6      -1.20      0.17    -1.53    -0.86 1.00     2455     3616
## wgP6M     -1.42      0.17    -1.74    -1.07 1.00     2552     3730
## wgPG       0.36      0.17     0.04     0.71 1.00     2486     3831
## wgPGG     -0.31      0.16    -0.62     0.03 1.00     2381     3981
## wgPM      -0.79      0.17    -1.11    -0.44 1.00     2546     3953
## wgPMG     -0.96      0.16    -1.27    -0.63 1.00     2452     3772
## wgPMM     -1.17      0.17    -1.49    -0.83 1.00     2311     3837
## 
## Family Specific Parameters: 
##       Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sigma     0.42      0.02     0.37     0.46 1.00    15742    14912
## 
## Samples were drawn using sampling(NUTS). For each parameter, Bulk_ESS
## and Tail_ESS are effective sample size measures, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).

2.2.3 EEG Control Data

We will also fit models to the control data. As we can see from Figure 2.4, the group differences are much smaller.

## Warning: Bulk Effective Samples Size (ESS) is too low, indicating posterior means and medians may be unreliable.
## Running the chains for more iterations may help. See
## http://mc-stan.org/misc/warnings.html#bulk-ess

Figure 2.3: The density plot shows the distribution of the empirical data, while the blue line shows the 66% and 95% prediction intervals.

3 Subgroup Comparisons

We will now compute the difference between sub- and super-groups.

3.1 Primary EEG Data

Finally, we calculate the probability that the RMS difference between subgroup and supergroup is larger than zero given the data. This information is then binned so we can colour in the posterior density plots.

Now we will use it!

Figure 3.1: Distributions of the difference in mean log(rms) between sub- and super-groups. The index of each relationship is indicated by the colour of the y-axis label. The fill of the density plots indicated the probability of the difference being greater than zero.

3.2 Psychophysical Data

We can do the same for the display duration thresholds from our psychophysics experiment. Here we are looking for the opposite effect, namely that display larger are larger for subgroups than for supergroups (see main paper), so we calculate the probability that differences in duration are smaller than zero.

Figure 3.2: Distributions of the difference in mean log display duration threshold between sub- and super-groups. The index of each relationship is indicated by the colour of the y-axis label. The fill of the density plots indicated the probability of the difference being less than zero.

3.3 Control EEG Data

We will now do exactly the same with the control data (odd harmonic data from parietal electrodes and even harmonic data from occipital electrodes)

Figure 3.3: Distributions of the difference in mean log(rms) between sub- and super-groups. The index of each relationship is indicated by the colour of the y-axis label. The fill of the density plots indicated the probability of the difference being greater than zero.

Figure 3.4: Distributions of the difference in mean log(rms) between sub- and super-groups. The index of each relationship is indicated by the colour of the y-axis label. The fill of the density plots indicated the probability of the difference being greater than zero.

3.4 Summary

We can summarise the subgroup comparison plots above by plotting ROC curves for each of our four measurements (Figure 3.5).

## # A tibble: 1 x 5
##     eeg threshold occ_even par_odd     p
##   <dbl>     <dbl>    <dbl>   <dbl> <dbl>
## 1    57        49       33      23  0.95

Figure 3.5: This figure shows how many of our 64 comparisons are classed as having a greater-than-zero difference (less-than-zero for the display durations) for difference thresholds. between 0.5 an 1.0.

If we take p=0.95 as our cut-off, we can see that the subgroup relations are preserved in 56/64 = 87.5% and 49/64 = 76.6% of the comparisons for the primary EEG and display durations repectively. This compares to the $32/64= 50.0% and 22/64 = 34.4% for the control EEG conditions.

4 Additional Analysis

4.1 Replicatation of Kohler et al (2016)

We can lok at the groups that only contain orientation, and see if we obtain the parametric response as documented in Kohler et al. (2016). We one might predict based on those results, we get a corresponding pattern of display duration thresholds, with the time taken to detect the symmetry dreasing as we increase the amount of rotational symmetry.

Curiosuly, we do not have a large difference betwee P4 and P6 in the eeg data.

Figure 4.1: The effect of increasing the amount of rotational symmetry.

4.2 Index and Normality

Subgroup relations can be classified by their index, and by whether they are normal or not. Here we investigate the extent to which these two variables can account for the variation between the subgroup relationships.

##  Family: gaussian 
##   Links: mu = identity; sigma = identity 
## Formula: mean_value ~ index * normal 
##    Data: comp_summary$eeg (Number of observations: 69) 
## Samples: 4 chains, each with iter = 2000; warmup = 1000; thin = 1;
##          total post-warmup samples = 4000
## 
## Population-Level Effects: 
##                  Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS
## Intercept            0.04      0.14    -0.24     0.32 1.00     2730
## index                0.06      0.02     0.02     0.11 1.00     2691
## normalTRUE           0.03      0.18    -0.31     0.38 1.00     2205
## index:normalTRUE     0.08      0.05    -0.02     0.17 1.00     2443
##                  Tail_ESS
## Intercept            3162
## index                2940
## normalTRUE           2308
## index:normalTRUE     2646
## 
## Family Specific Parameters: 
##       Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sigma     0.28      0.03     0.24     0.34 1.00     3221     2398
## 
## Samples were drawn using sampling(NUTS). For each parameter, Bulk_ESS
## and Tail_ESS are effective sample size measures, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).

##  Family: gaussian 
##   Links: mu = identity; sigma = identity 
## Formula: mean_value ~ index * normal 
##    Data: comp_summary$threshold (Number of observations: 69) 
## Samples: 4 chains, each with iter = 2000; warmup = 1000; thin = 1;
##          total post-warmup samples = 4000
## 
## Population-Level Effects: 
##                  Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS
## Intercept           -0.47      0.24    -0.95    -0.00 1.00     2759
## index               -0.07      0.04    -0.15     0.01 1.00     2714
## normalTRUE           0.45      0.31    -0.16     1.04 1.00     2245
## index:normalTRUE    -0.14      0.08    -0.30     0.02 1.00     2326
##                  Tail_ESS
## Intercept            2846
## index                2830
## normalTRUE           2407
## index:normalTRUE     2307
## 
## Family Specific Parameters: 
##       Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
## sigma     0.46      0.04     0.39     0.55 1.00     3146     2747
## 
## Samples were drawn using sampling(NUTS). For each parameter, Bulk_ESS
## and Tail_ESS are effective sample size measures, and Rhat is the potential
## scale reduction factor on split chains (at convergence, Rhat = 1).

We can see that the index of the subgroup relationship has an effect on both the difference in log(rms) and the difference in log(display duration): relationships with a higher index lead to larger differences.

Figure 4.2: The effect of index and normality on log(rms) and log(ms).

4.3 Correlation Between Primary EEG data and Psychophysical Thresholds

Finally, we will investigate whether there is a correlation between the our primary EEG measure (rms amplitude of odd harmonics over occipital cortex) and our display duration thresholds. As our two different measures come from different samples of participants, we are unable to do a direct comparison. However, we can use the results of the models discussed in Section 3 and check for a correlation between the predicted values of the two measures.

##     Estimate  Est.Error      Q2.5     Q97.5
## R2 0.4789912 0.06662853 0.3319635 0.5860021

We can see that although the correlation is relatively weak, our confidence interval indicates that we can be reasonably positive that \(R^2>0\) (i.e, 95% credible interval is 0.026 - 0.302).

Figure 4.3: Scatter plot showing the correlation between our two measures. Each line is a sample from the posterior of a Bayesian linear regression.